With reference to training our own model, I would like to ask if it is possible to use transfer learning and initialise the model using pre-trained weights (e.g. from a model trained with ImageNet), rather than training from randomly initialised weights. I have tried doing this using retrain_from, but since the number of outputs are different, it does not work. Also wondering if the argument transfer_mode has anything to do with this.
Hi @mikeyEcology,
With reference to training our own model, I would like to ask if it is possible to use transfer learning and initialise the model using pre-trained weights (e.g. from a model trained with ImageNet), rather than training from randomly initialised weights. I have tried doing this using
retrain_from, but since the number of outputs are different, it does not work. Also wondering if the argumenttransfer_modehas anything to do with this.Thank you.
Best regards, Joejyn